Abstract Channel estimation is crucial to achieving wide-area coverage for ultra-low-cost and low-power narrowband Internet of Things (NB-IoT) devices that are in coverage extremities. Radio coverage can be extended by repeatedly transmitting the same signal over a protracted period. In repetition dominated NB-IoT systems, existing channel estimators extensively used in the orthogonal frequency-division multiplexing (OFDM) system may be no longer applicable due to their considerable computational complexity and power consumption. In this paper, we propose narrowband demodulation reference signal (NDMRS)-assisted transform-domain low-complexity channel estimation algorithms named random sorting least squares (RS-LS), and de-noising LS (D-LS). Another sub-optimal estimator, stemming from the filtered channel estimates called linear minimum mean square error-approximation (LMMSE-A) is also studied. We first estimate initial channel response at pilot frequencies using the conventional LS method; and then, apply several additional operations in time-domain to suppress LS estimation error without exploiting extra frequency-band resources, and increasing significant computational complexity. Finally, channel estimates for the remaining OFDM symbols within an NB-IoT subframe are obtained by employing the time dimensional linear interpolation. Through several simulation examples, the viability of the proposed estimators is verified in comparison with the conventional LS, denoise, and optimal LMMSE estimators in terms of channel mean square error (MSE), block error rate (BLER), and throughput against signal-to-noise ratio (SNR) for Long Term Evolution (LTE)-based uplink NB-IoT systems.